BACKGROUND INFORMATION

[0001]
1. Field of Invention

[0002]
Inventionrelates to masks, also known as photomasks, used in photolithography processes and, more particularly, to a method for applying levelset technology to timeevolve contours representing patterns on photomasks in such a way so as to allow for production of wafer patterns with minimal distortions or artifacts and to allow for the ability to constrain resulting contours to rectilinear patterns.

[0003]
2. Description of Related Art

[0004]
Lithography processing represents an essential technology for manufacturing Integrated Circuits (IC) and Micro ElectroMechanical Systems (MEMS). Lithographic techniques are used to define patterns, geometries, features, shapes, et al (“patterns”) onto an integrated circuit die or semiconductor wafer or chips where the patterns are typically defined by a set of contours, lines, boundaries, edges, curves, et al (“contours”), which generally surround, enclose, and/or define the boundary of the various regions which constitute a pattern.

[0005]
Demand for increased density of features on dies and wafers has resulted in the design of circuits with decreasing minimum dimensions. However, due to the wave nature of light, as dimensions approach sizes comparable to the wavelength of the light used in the photolithography process, the resulting wafer patterns deviate from the corresponding photomask patterns and are accompanied by unwanted distortions and artifacts.

[0006]
Techniques such as Optical Proximity Correction (OPC) attempt to solve this problem by appropriate predistortion of the photomask pattern. However, such approaches do not consider the full spectrum of possible photomask patterns, and therefore result in suboptimal designs. The resulting patterns may not print correctly at all, or may not print robustly. Accordingly, there is a need for a method for generating the optimal photomask patterns which result in the robust production of wafer patterns faithful to their target patterns.
SUMMARY OF INVENTION

[0007]
Photomask patterns are represented using contours defined by levelset functions. Given target pattern, contours are optimized such that defined photomask, when used in photolithographic process, prints wafer pattern faithful to target pattern.

[0008]
Optimization utilizes “merit function” for encoding aspects of photolithographic process, preferences relating to resulting pattern (e.g. restriction to rectilinear patterns), robustness against process variations, as well as restrictions imposed relating to practical and economic manufacturability of photomasks.
BRIEF DESCRIPTION OF DRAWINGS

[0009]
FIG. 1 is a diagram illustrating a simple example target pattern to be printed on a wafer using a photolithography process, according to an embodiment of the present invention.

[0010]
FIG. 2 is a diagram illustrating an example photomask pattern in the (x, y) plane comprising regions, according to a preferred embodiment of the present invention.

[0011]
FIG. 3 is a diagram showing an example wafer pattern illustrative of what might print on a wafer using the example photomask pattern of FIG. 2 in a photolithography process, according to an embodiment of the present invention.

[0012]
FIG. 4 a is a diagram illustrating a levelset function representing the example photomask pattern of FIG. 2 by defining the contours which enclose the regions in the photomask pattern, according to an embodiment of the present invention.

[0013]
FIG. 4 b is a diagram illustrating the levelset function of FIG. 4 a intersected with the zero plane parallel to the (x,y) plane, according to an embodiment of the present invention.

[0014]
FIG. 5 is a flow chart illustrating a method for timeevolving contours of a photomask pattern in order to minimize a Hamiltonian function, according to a preferred embodiment of the present invention.

[0015]
FIG. 6 a is a diagram illustrating a photomask pattern, according to an embodiment of the present invention.

[0016]
FIG. 6 b is a diagram illustrating a photomask pattern corresponding to a final levelset function output by the algorithm, according to an embodiment of the present invention.

[0017]
FIG. 6 c is a diagram illustrating a wafer pattern as produced using the photomask pattern of FIG. 6 b in a photolithography process.

[0018]
FIG. 6 d is a diagram illustrating a rectilinear photomask pattern output by the algorithm based on the initial photomask shown in FIG. 2, according to an embodiment of the present invention.

[0019]
FIG. 6 e is a diagram illustrating a wafer pattern as produced using the rectilinear photomask pattern of FIG. 6 d in a photolithography process, according to an embodiment of the present invention.

[0020]
FIG. 7 is a diagram illustrating a 2dimensional subspace of an mdimensional solution space of levelset functions, showing Hamiltonian H as a function of ψ(x_{1},y_{1}) and ψ(x_{2},y_{2}), according to an embodiment of the present invention.
DETAILED DESCRIPTION OF PREFERRED EMBODIMENT(S)

[0021]
As understood herein, the term “wafer pattern” is understood to include any polygon (rectilinear or nonrectilinear) or other shape or pattern to be formed on a semiconductor or other material substrate, for example digital or analog circuit structures or interconnect.

[0022]
FIG. 1 is a diagram illustrating an example target pattern 100 to be printed on a wafer using a photolithography process. Target pattern 100 comprises regions 101 enclosed by contours 102. Preferably, areas within regions 101 represent photoresist and the area outside regions 101 represents the absence of photoresist.

[0023]
FIG. 2 is a diagram illustrating an example photomask pattern 201 in the (x, y) plane comprising regions 201 for printing a wafer pattern in a photolithography process. In a preferred embodiment, an area within a region 201 represents chrome and the area outside regions 201 represents glass on the photomask. Alternatively, an area within a region 201 represents a material other than chrome and the area outside regions 201 represents a material other than glass on the photomask.

[0024]
FIG. 3 is a diagram showing an example wafer pattern 300 illustrative of what might print on a wafer using photomask pattern 200 in a photolithography process. Preferably, areas within regions 301 represent photoresist and the area outside regions 301 represents the absence of photoresist. Note that wafer pattern 300 differs from target pattern 100 due to distortions and artifacts produced by the photolithography process. It is an object of the present invention to generate a photomask pattern which, when used in a photolithography process, produces a wafer pattern more faithful to the corresponding target pattern, the wafer pattern having fewer undesirable distortions and artifacts.

[0025]
Because we use contours to define regions in a photomask pattern, we use a mathematical description of such contours.
FIG. 4 a illustrates a levelset function ψ(x, y)
400 representing example photomask pattern
200 by defining the contours which enclose the regions in photomask pattern
200. ψ(x, y) is a function with the property that

 1. ψ(x,y)=0 everywhere along the boundary of a region;
 2. ψ(x, y)>0 “inside” a region (for example, those regions corresponding to the chrome portions of the mask);
 3. ψ(x, y)<0, or is negative “outside” a region (for example, those regions corresponding to the clear quartz portions of the mask).
The contours are defined by the “levelset”, i.e. those values in the (x,y) plane such that ψ(x,y)=0. FIG. 4 b illustrates the levelset by intersecting the levelset function 400 with the zero plane 401 parallel to the (x,y) plane.

[0029]
It is an aspect of the present invention to, given a target pattern, find a levelset function ψ(x, y) such that the levelset ψ(x,y)=0 defines a set of contours, which, when interpreted as the boundaries of regions of a pattern on a photomask, correspond to the design of a photomask producing a wafer pattern with little distortions and artifacts compared to the target pattern, wherein the wafer pattern results from a photolithography process using the photomask. The extent to which the set of contours defined by a levelset function ψ(x, y) is optimal is calculated using a functional known as the “merit function”, also referred to herein as the “Hamiltonian” H. The Hamiltonian H of a levelset function ψ(x, y) is indicative of the degree of similarity between the printed wafer pattern and the desired target pattern, the printed wafer pattern resulting from a photolithography process using a photomask given by the contours defined by ψ(x, y). (We call the “merit function” the “Hamiltonian” by way of analogy to the Hamiltonian function used in classical dynamics or quantum mechanics).

[0030]
Mathematically, the Hamiltonian H is a functional which maps a function to a real number:
H:C(R^{2})→R
Optionally, the Hamiltonian also depends upon a number of real parameters, or, as described below, is a function of multiple levelsets. The Hamiltonian is chosen so that the optimal solution has the smallest value, i.e. we seek to find a levelset function which minimizes the Hamiltonian. It follows that, once an appropriate Hamiltonian is specified, the problem of finding an optimally designed photomask is equivalent to the problem of finding a levelset function which minimizes the Hamiltonian. It also follows that the specification of an appropriate Hamiltonian is a valuable step in applying the principles of our invention, given that the form of the Hamiltonian functional will directly determine the contours which result from the optimization problem.

[0031]
Note that our description of the problem in terms of minimizing a Hamiltonian function is for purposes of description and not by way of limitation. Equivalent alternatives are available to one of ordinary skill in the art based on the presented description (for example formulating the optimization problem as a maximization problem rather than a minimization problem, or assigning negative values to the levelset function values at points on the inside of enclosed regions and assigning positive values to the points on the outside of enclosed regions, or by choosing a levelset other than the zero levelset to specify contours, etc.).

[0032]
FIG. 5 is a flow chart illustrating a method for timeevolving contours of a photomask pattern in order to minimize a given Hamiltonian function, according to a preferred embodiment of the present invention. FIG. 5 depicts the steps used to find a levelset function which defines an optimal photomask for a given target pattern. The levelset function is found by iteratively refining an initial guess so that the refinements progressively result in better “merit” values, i.e. decrease a Hamiltonian H, wherein H depends on the target pattern, the particular photolithography process under consideration, constraints on the photomask manufacturing, and other factors to be described in detail below. We briefly describe the steps in FIG. 5 prior to providing the detail.

[0033]
Start 501 with a set of initial inputs, which may include a target pattern, parameters describing the particular photolithography process under consideration, mask manufacturing restrictions, and other factors to be described in detail below. Initialize 502 i=0 and choose 503 an initial levelset function ψ(x, y)=ψ(x, y). Determine 504 whether ψ_{i}(x, y) is acceptable (details on determining this below). If ψ_{i}(x, y) is 505 determined to be acceptable, output 506 ψ_{i}(x, y) as the result of the minimization, and finish. Otherwise 507, increment 508 i by one and choose 509 a next ψ_{i}(x, y) so as to gain an improvement (details on choosing next ψ_{i }appear below), repeating until ψ_{i}(x, y) is determined 507 to have acceptable “merit”, and finishing by outputting 506 the final ψ_{i}(x, y) as the result of the minimization. Because the initial levelset function ψ_{0 }changes through each iteration, it can be thought of as evolving with time, and it is convenient to think of each successive function ψ_{i}(x, y) as a discrete snapshot of a continuously evolving “timedependent levelset function” ψ(x,y,t) of space and time (t denoting time).

[0034]
FIG. 6 a illustrates a photomask pattern 602 corresponding to a levelset function ψ_{i}(x, y) after about 500 iterations of the algorithm. FIG. 6 b illustrates a photomask pattern 603 corresponding to a final levelset function output by above optimization algorithm. FIG. 6 c illustrates a wafer pattern 604 as produced using photomask pattern 603 of FIG. 6 b in a photolithography process.

[0035]
In one embodiment, a succeeding function ψ_{i+1}(x, y) is chosen by adding a small change Δ_{i}(x, y) to the current levelset function ψ_{i}(x, y), wherein Δ_{i}(x, y) is another function over the same domain as ψ_{i}(x, y):
ψ_{i+1}(x,y)=ψ_{i}(x,y)+Δ_{i}(x,y) (1)
In a preferred embodiment, a succeeding function ψ_{i+1}(x, y) is chosen by first adding a small change Δ_{i}(x, y) to the current levelset function ψ_{i}(x, y) and then projecting the resulting sum onto a subspace which constrains the contours to be rectilinear (details on the projection appear below). FIG. 6 d illustrates a rectilinear photomask pattern 605 output by above algorithm based on the initial photomask 200 shown in FIG. 2. FIG. 6 e illustrates a wafer pattern 606 as produced using rectilinear photomask pattern 605 of FIG. 6 d in a photolithography process.
In one embodiment of our invention, we calculate Δ_{i}(x, y) as follows:
$\begin{array}{cc}{\Delta}_{i}\left(x,y\right)=\Delta \text{\hspace{1em}}t\xb7\left\{\frac{\delta \text{\hspace{1em}}H}{\mathrm{\delta \psi}}{}_{\psi ={\psi}_{i}}+R\left({\psi}_{i}\right)\right\}\xb7\uf603\nabla {\psi}_{i}\uf604& \left(2\right)\end{array}$
where Δt is a small constant hereinafter referred to as “timestep”,
$\frac{\delta \text{\hspace{1em}}H}{\mathrm{\delta \psi}}$
is the Frechet derivative of the Hamiltonian H, R(ψ) is a “regularization term” which slightly modifies the evolution of the levelset function to improve numerical stability, and ∇ψ_{i} is the norm of the gradient of the levelset function ψ_{i}(x, y). Each of these terms as well as the projection operation will be described in more detail below.

[0036]
In still another embodiment of our invention, we use the continuoustime version of equation (2) above and timeevolve the timedependent levelset function according to the equation
$\begin{array}{cc}\frac{\partial}{\partial t}\psi \left(x,y,t\right)=\left\{\frac{\delta \text{\hspace{1em}}H}{\mathrm{\delta \psi}}+R\left({\psi}_{i}\right)\right\}\xb7\uf603\nabla \psi \uf604& \left(3\right)\end{array}$
which can be implemented computationally using a variety of techniques different from the discretization described in equation (2) above, but that are known to one of ordinary skill in the art.

[0037]
In a preferred embodiment, and to facilitate computation, a levelset function ψ_{i}(x, y) is represented by a discrete set of m function values over a set of m points in the (x, y) plane. In one embodiment, the set of m points comprises a grid spanning an area representing the photomask. Alternatively, the set of m points is chosen according to a different arrangement in the area representing the photomask. From this perspective, a levelset function ψ_{i}(x, y) and a “small change” function Δ_{i}(x, y) are determined by their values at the set of m points and consequently can be considered as m dimensional vectors in “solution space.” FIG. 7 is an illustration of the possible values for the first two components of an mdimensional vector representing a levelset function, i.e., illustrating a 2dimensional subspace of solution space. In the subspace shown, we plot H as a function of ψ(x_{1},y_{1}) and ψ(x_{2},y_{2}). For this example, both ψ(x_{1},y_{1}) and ψ(x_{2},y_{2}) can vary between −1 and +1. The minimum in this example is seen to occur at ψ(x_{1},y_{1})=0.3 and ψ(x_{2},y_{2})=−0.2. Starting with an initial guess at a levelset function ψ_{0}(x, y), we approach the minimum by taking a small step (in step 509) in the direction of “steepest descent” to obtain a new location which is closer to a minimum. By repeating this process we quickly reach a minimum. Timeevolving the levelset function according to above preferred embodiment is analogous to the foregoing, except that the dimensionality of the entire “solution space” is much greater than 2 (for example, it can equal the number of grid points m in the discretized version, or be infinite in a continuous version).

[0038]
From the above discussion, it is seen that one can find the minimum without actually calculating the Hamiltonian. However, it may be useful to calculate the Hamiltonian in order to determine the “merit” of the current levelset function. For example, it may be reasonable to stop iterating, even before the algorithm converges on an optimal solution, if a sufficient solution has been found. Similarly, one may wish to check the Hamiltonian occasionally (every several iterations), or only at those times when an adequate solution seems likely to have been found, such as when the levelset evolution generates only small changes in the succeeding levelsets. At this point we shall reexamine the steps of the flow chart of FIG. 5 in more detail:

[0000]
Input

[0039]
The algorithm starts
501 with a set of inputs, among them a target pattern given in a particular format (“pattern I/O formats”). The target pattern may be presented in a variety of formats, including, but not limited to:

 1. Image formats, such as bitmaps, JPEGs (Joint Photographic Experts Group), or other image formats;
 2. Semiconductor industry formats, such as CIF, GDSII, Oasis, OpenAccess; or
 3. Proprietary formats, such as an Electronic Design Automation (EDA) layout database.
The target pattern itself may be a representation of various types of content, for example (but not limited to):
 1. One or more levels of an IC design for a particular IC type;
 2. A pattern for a nonIC application (e.g. a MEMs device);
 3. A pattern which can be used as part of a larger design, such as a standard cell, or DRAM bit cell, etc.
The algorithm also accepts one or more constraints as input, including (but not limited to) target pattern or mask pattern constraints, which may be specified as rules (such as critical dimensions, tolerances, etc.); and target pattern or mask pattern constraints specified as additional images (in a “pattern I/O format”) specifying for example maximal or minimal areas to be covered, or critical regions, etc.

[0046]
It is contemplated that the teachings of the present invention can be used to refine a photomask pattern determined through some other process, and that the output of the algorithm of present invention could be fed, or otherwise used, as input into another technique or methodology for optimally providing a photomask. In general, an iterative process is contemplated in which a mask pattern is taken through a series of transformations, with the teachings of the present invention accomplishing a subset of those transformations.

[0047]
Other accepted inputs include parameters of the Hamiltonian function, including but not limited to parameters of the physical model used to simulate the photolithographic process, and parameters indicating the desired attributes of the eventual solution. These may include, for example, the number and type of photomasks employed, the wavelength of a stepper machine, the type and properties of the illumination, the type and properties of the photoresist, the type and properties of the lens, etc. Other parameters may include properties of error sources, such as defocus, exposure, alignment, defects, etc.

[0000]
Initialization

[0048]
After receiving inputs in step 501 and initializing 502 i=0, we initialize 503 the levelset function ψ_{0}. In theory, almost any levelset initialization should be sufficient; however, initial conditions can have an impact on the time required for convergence and therefore on the cost of the process. Moreover, it is possible that for sufficiently poor initial conditions the algorithm will fail to converge.

[0049]
A variety of ways to initialize the levelset function will be apparent to one of ordinary skill in the art. In one embodiment of the present invention, the initial levelset function is chosen, according to an initial photomask pattern comprising enclosed regions (to be chosen below), by assigning

 1. the value +1 to ψ_{0}(x, y) everywhere within the enclosed regions of the photomask pattern;
 2. the value −1 to ψ_{0}(x, y) everywhere outside the enclosed regions of the photomask pattern; and
 3. the value 0 to ψ_{0}(x, y) at the boundaries (contours) of the regions of the photomask pattern.

[0053]
However, it is desirable to have a smoother and approximately continuous function as the levelset function. In a preferred embodiment of the present invention, the levelset function is a “distance function,” wherein the value of the function at a given point represents the (signed) distance of the point to the (nearest) boundary of a region in photomask pattern (positive inside a region's boundary, negative outside). Such a distance function has a variety of useful properties. For example, in the context of the present invention, a distance function allows for computations that depend not just on what is inside a region's boundary or outside a region's boundary, but what is “near” the boundary, where “near” is functionally based on distance. As the levelset function evolves, it slowly loses its property as a distance function. However, this can be corrected using a “redistancing” process, known to one of ordinary skill in the art.

[0054]
It remains to determine an initial photomask pattern on which to base the choice of the initial levelset function ψ
_{0}(x, y) in step
503. A variety of possible choices are available including (but not limited to) the following:

 1. Random. This is unlikely to be the choice leading to fastest minimization, but it should be very robust;
 2. The target pattern. Especially for the case of a single chrome and glass mask, choosing the initial mask pattern to be equal to the target pattern is likely to perform fairly well. This is because it is possible for the final mask pattern to be similar to the target pattern;
 3. The result of heuristics applied to the target pattern. For example, an OPC algorithm is applied to the target pattern, with the result used as the initial mask pattern. For multiple mask processes, one approach is to use heuristics to split the pattern into multiple exposures, for example, separating horizontal and vertical lines;
 4. Results from previous solutions to the same or similar problems. These are likely to be similar to the desired final pattern; or
 5. Results from solutions to other similar regions on the mask. As above, these are likely to yield similar solutions. One can imagine, for example, that the mask comprises individual mask areas. As pattern on such individual areas are optimized, the solutions can be used as initial guesses for other areas. Since the optimal solution for a given area will depend upon interactions with neighboring areas, the solutions may not be the same. However, the optimal solution from one area may serve as a good initial guess for another similar area.

[0060]
There are numerous ways to initialize an original target photomask pattern. The previous possibilities that have been described are meant only as a partial list of possible alternatives.

[0061]
In one embodiment, a levelset function is stored as an array of values representing the value of the levelset function at fixed points on a 2dimensional grid. Optionally, a more sophisticated approach (referred to as “local levelset”) only stores the values near the boundaries; depending upon the pattern and the resolution, this can be significantly more efficient. Other ways of representing and storing a levelset function will be apparent to one of ordinary skill in the art.

[0000]
Checking the “Merit”

[0062]
As seen in the flow chart, in step 504 the algorithm determines if it has converged upon a suitable set of contours so as to provide an optimal photomask. In one embodiment, this check is performed after each step in the evolution of the contours (defined by the levelsets). Alternatively, this check is performed after two or more steps of the levelset evolution.

[0063]
One simple method to determine whether an acceptable solution has been found (in step 504) is to calculate the value of the Hamiltonian H(ψ_{i}) resulting in a “merit” of the current levelset solution. Alternatively, a solution is deemed acceptable based on the number of iterations performed. It may be advantageous to use locally defined criteria to stop iterating in areas of the photomask where the solution is already acceptable, and to continue iterating in areas where the solution has not yet reached an acceptable level of merit. In this context, “local” can mean on the level of a pixel, on the level of a small area (for example, an interaction distance), on the level of a hierarchical subdivision of the mask area, or on other alternative levels.

[0064]
Yet another indication of convergence is provided by the magnitude of the gradient (in “solution space”) or Frechet derivative—as the contours approach an optimal state, the derivative decreases and approaches zero. Similarly, the change in the shape of the contours from one iteration to another iteration provides an indicator of convergence. Although we have described several indicators, one of ordinary skill in the art will recognize other indicators as well.

[0000]
TimeEvolving Contours

[0065]
As described above, in a preferred embodiment a levelset function evolves in a series of steps in which we add to it a small function Δ
_{n}(x, y) calculated via equation (2)
${\Delta}_{i}\left(x,y\right)=\Delta \text{\hspace{1em}}t\xb7\left\{\frac{\delta \text{\hspace{1em}}H}{\mathrm{\delta \psi}}{}_{\psi ={\psi}_{i}}+R\left({\psi}_{i}\right)\right\}\xb7\uf603\nabla {\psi}_{i}\uf604$
It is common for the optimization problem of the present invention to be mathematically “illposed” in the sense that the solution may be nonunique. In order to avoid inherent numerical instabilities during the time evolution we employ a “regularization” technique, adding a small term R(ψ) to the Hamiltonian H in order to help stabilize the time evolution. The resulting contours will have less “noise” and appear smoother to the eye. There are many ways to add regularization which will be apparent to those skilled in the art, including (but not limited to):
$1.\text{\hspace{1em}}R\left(\psi \right)=\nabla \xb7\frac{\nabla \psi}{\uf603\nabla \psi \uf604}$

 Mean curvature regularization—Adding this term tends to reduce noise in the image by minimizing the length of contours.
$2.\text{\hspace{1em}}R\left(\psi \right)=\nabla \xb7\frac{\nabla \psi}{\uf603\nabla \psi \uf604}\stackrel{\_}{\nabla \xb7\frac{\nabla \psi}{\uf603\nabla \psi \uf604}}\text{\hspace{1em}}\left(\mathrm{with}\text{\hspace{1em}}\mathrm{the}\text{\hspace{1em}}\mathrm{bar}\text{\hspace{1em}}\mathrm{indicating}\text{\hspace{1em}}\mathrm{average}\right).$
 Average mean curvature—This tends to minimize the length of the boundaries while keeping the total area of the enclosed regions fixed, giving preference to smoother contours and contours enclosing larger regions, since larger regions have less boundary per unit area as compared to many small regions.
$3.\text{\hspace{1em}}R\left(\psi \right)=\frac{\partial}{\partial x}\left(\frac{{\psi}_{x}}{\uf603{\psi}_{x}\uf604}\right)+\frac{\partial}{\partial y}\left(\frac{{\psi}_{y}}{\uf603{\psi}_{y}\uf604}\right)$
 Wulfcrystal regularization or Wulf curvature—This is similar to curvature except that it prefers Manhattan geometries. Other variations of Wulf regularization can be used preferring straight edges in Manhattan geometries or 45 degree angles. Use of Wulfcrystal regularization may be helpful in the design of masks with rectilinear contours, although it will not guarantee rectilinear geometries.
$4.\text{\hspace{1em}}R\left(\psi \right)=\frac{\partial}{\partial x}\left(\frac{{\psi}_{x}}{\uf603{\psi}_{x}\uf604}\right)+\frac{\partial}{\partial y}\left(\frac{{\psi}_{y}}{\uf603{\psi}_{y}\uf604}\right)\stackrel{\_}{\frac{\partial}{\partial x}\left(\frac{{\psi}_{x}}{\uf603{\psi}_{x}\uf604}\right)+\frac{\partial}{\partial y}\left(\frac{{\psi}_{y}}{\uf603{\psi}_{y}\uf604}\right)}$

[0069]
Average Wulf curvature—Combining aspects of average mean curvature and Wulf curvature, this exhibits a preference for rectilinear contours enclosing large regions.

[0070]
In all of the above regularization expressions, it is possible for the denominator in one or more of the fractions to equal zero. In order to avoid dividing by zero, one can add a small constant to the denominator, or set the expression equal to zero whenever both the numerator and denominator equal zero.

[0071]
One of ordinary skill in the art will recognize other possibilities for regularization. It should be obvious that it may be desirable for the amount or type of regularization to change with time as the contours evolve, and that alternative ways of introducing regularization are apparent to those skilled in the art and are part of the teachings of the present invention.

[0072]
In equation (2), as well as in several of the regularization expressions, we need to compute ∇ψ. The way in which the gradient is computed can have important consequences in terms of numerical stability. Preferably, techniques for calculating gradients known to those skilled in the art as HamiltonJacobi Essentially NonOscillatory (ENO) schemes or HamiltonJacobiWeighted Essentially NonOscillatory (WENO) schemes are used. Alternatively, other ways of computing the gradient are used, as are known to one of ordinary skill in the art.

[0073]
In a similar vein, the time evolution of the timedependent levelset function can be implemented using a variety of numerical techniques. One embodiment described above uses what is known as “first order Runge Kutta”. Alternatively, other variations such as third order Runge Kutta can be used as will be obvious to those skilled in the art.

[0074]
The method of gradient descent involves multiple iterations; the size of the function Δ_{i}(x, y) chosen as part of performing step 509 is scaled by the “timestep” Δt appearing in equation (2). The larger the timestep, the faster the system converges, as long as the timestep is not so large so as to step over the minimum or lead to numerical instabilities. The convergence speed of the algorithm can be improved by choosing an appropriate timestep.

[0075]
There are numerous ways to choose the timestep Δt. In one embodiment, we choose a time step that is just small enough so as to guarantee that the system is stable. In an alternative embodiment, we start with a large time step and gradually reduce the time step as the algorithm approaches a minimum. In an alternative embodiment, we vary the time step locally and per subarea of the photomask. Other approaches will be known to those skilled in the art, or other means of adapting the time step to the specific situation.

[0076]
In another embodiment, one can use what are known as implicit methods, optionally with linearpreconditioning, in order to allow for a larger timestep. Still other variations will be known to one of ordinary skill in the art.

[0077]
Analogous to refining the time granularity by reducing the timestep, in one embodiment of the present invention the granularity or placement of the grid of points on the photomask is adjusted in a timedependent fashion as the algorithm approaches convergence. By performing the initial iterations on a larger grid, and increasing the number of grid points with time as greater accuracy is desired, a solution is obtained more quickly. Other such “multigrid” techniques will be apparent to those skilled in the art. Another possibility is using an adaptive mesh technique, whereby the grid size varies locally.

[0078]
It is possible that the process of timeevolving the contours arrives at a configuration from which there is no “downhill” path in “solutionspace” to a solution, or in which such paths are inordinately long or circuitous. In such a state, convergence may require a large number of iterations. Also, the algorithm may get “stuck” at a local (but not global) minimum. Some example techniques for handling such a situation are as follows:

 1. Changing the Hamiltonian. Various modifications can be made to the Hamiltonian in order to bridge local minima in solution space; for example, regularization terms can sometimes be used for this purpose;
 2. Addition of random bubbles. Adding random noise to a levelset function will create new regions, which can then timeevolve into the solution. Noise (that is distortion) can be purposefully added at random or it can be targeted in specific regions (for example, known problematic target geometries, or regions which are not converging on their own to acceptable errors, etc.);
 3. Heuristic bubbles. Rather than adding random noise, a specific modification feature, known from experience to generally help the system converge, is added; for example, if certain areas appear to be evolving too slowly, one could add a constant to the levelset function in that area, thereby making all the features in that area “bigger”;
 4. Uphill steps. By making uphill moves, either at random, or specifically in places, the algorithm avoids getting stuck in local minima and works toward a global minimum. Similar techniques from discrete optimization or simulated annealing which are useful to the algorithm of the present invention will be apparent to one of ordinary skill in the art.
Alternatives to the previous example techniques will be apparent to those of ordinary skill in the art.
Projection Operator

[0083]
In many cases, it is desired to constrain the solution to rectilinear contours, for example to improve manufacturability or reduce costs of the corresponding masks. The present invention enforces this constraint by using a projection operator.

[0084]
The projection operator can be understood by considering the solutionspace of all possible contours, and recognizing that the set of rectilinear contours is a subspace of the solutionspace. The projection operator constrains the evolution of the contours to within the subspace of rectilinear contours.

[0085]
The projection operator takes a set of possibly curvilinear contours and approximates them with a set of rectilinear contours. In one embodiment, choose a fixed grid size (possibly corresponding to manufacturing capabilities), and “round” every contour to the nearest grid. This is accomplished, for example, by setting the levelset function to be positive if the majority of the points within a single gridcell are positive, negative if the majority of the points within a single gridcell are negative, and zero along the boundaries. Alternative implementations of the projection operator will be apparent to one of ordinary skill in the art.

[0086]
In one embodiment of the present invention, the projection operator is applied to the levelset function after each timestep iteration. In this way, the contours are always constrained to be rectilinear. In an alternative embodiment, the contours are allowed to evolve for multiple timesteps between applications of the projection operator, in which case the contours may temporarily deviate from strict rectilinear form. The preferred frequency of projection will depend upon factors such as speed of computation, choice (or implementation) of the projection operator, or the manner (or implementation) of the curvilinear timeevolution.

[0000]
Merit Function/Hamiltonian

[0087]
As illustrated, the optimization problem and the resulting contours are determined by a merit function referred to as the Hamiltonian. There are many alternative ways to choose a merit function within the scope of the present invention. In one embodiment, the Hamiltonian comprises a sum of two parts:

 1. A first part, based upon the levelset contours themselves; and
 2. A second part, based upon the resulting pattern which would be printed on a wafer or die given the photomask corresponding to the levelset contours.

[0090]
The first part of the Hamiltonian may comprise one or more terms such that the resulting optimized photomask pattern has properties which are desirable from a manufacturing point of view; for example, the “regularization” terms described above can be viewed as elements of the Hamiltonian that exhibit a preference for contours corresponding to more easily manufacturable masks.

[0091]
The second part of the Hamiltonian takes into consideration a model of the photolithographic process, i.e. a method for calculating the wafer pattern resulting from a particular mask pattern (a “forward model”). Following describes an example forward model for one embodiment of the present invention.

[0092]
In a typical photolithographic process, light passes through the photomask and the lens, and then falls upon the wafer, where the photoresist is exposed. For coherent illumination, the electric field falling upon the photomask is approximately constant. The clear regions of the mask pass the light, while the opaque regions block the light. It follows that the electric field, just behind the mask, can be written as:
$M\left(\overrightarrow{r}\right)=\left\{\begin{array}{cc}0& \mathrm{chrome}\\ 1& \mathrm{glass}\end{array}\right\}$
where {right arrow over (r)}=(x,y) is a point on the (x,y) plane. Corresponding to an embodiment of the present invention wherein the regions in which the levelset function is positive indicate glass and the regions in which the levelset function is negative indicate chrome (with the levelset equal to zero at the boundaries or contours), one can write the previous expression as a function of a levelset function ψ(x,y) as follows:
M({right arrow over (r)})=ĥ(ψ(x,y))
wherein ĥ is the Heaviside function:
$\hat{h}\left(x\right)=\left\{\begin{array}{cc}1& x\ge 0\\ 0& x<0\end{array}\right\}$
Because an ideal diffraction limited lens acts as a lowpass filter, this will serve as a good approximation to the actual (almost but not quite perfect) lens used in a typical photolithographic process. Mathematically, the action of the lens would then be written as follows:
A({right arrow over (r)})=f ^{−1}({circumflex over (C)}(f(M({right arrow over (r)}))))
where A({right arrow over (r)}) indicates the electric field distribution on the wafer, f indicates a Fourier transform, f^{−1 }indicates an inverse Fourier transform, and Ĉ indicates the pupil cutoff function, which is zero for frequencies larger than a threshold determined by the numerical aperture of the lens, and one otherwise:
$\stackrel{\u22d2}{C}\left({k}_{x},{k}_{y}\right)=\hat{h}\left({k}_{\mathrm{max}}^{2}\left[{k}_{x}^{2}+{k}_{y}^{2}\right]\right)=\left\{\begin{array}{c}0\text{\hspace{1em}}{k}_{x}^{2}+{k}_{y}^{2}\ge {k}_{\mathrm{max}}^{2}\\ 1\text{\hspace{1em}}{k}_{x}^{2}+{k}_{y}^{2}<{k}_{\mathrm{max}}^{2}\end{array}\right\}$
wherein k_{x}, k_{y}, and k_{max }represent frequency coordinates in Fourier space.
Finally, we determine the image in the photoresist upon the wafer. In one embodiment this process is modeled using a “threshold resist”: in regions where the intensity is greater than a given threshold (which we shall call I_{th}), the resist is considered exposed; in regions below the threshold, the resist is considered unexposed. Mathematically, this is handled once again with a Heaviside function:
I({right arrow over (r)})=ĥ(A({right arrow over (r)})^{2} −I _{th})
Combining the above, we find that:
F(ψ(x,y))={circumflex over (h)}(f−1 ({circumflex over (C)}(f({circumflex over (h)}(ψx,y)))))^{2} −I _{th }
This is a selfcontained formula which reveals the wafer pattern corresponding to the photomask pattern defined by the levelset function, within the context of the model just described. It should be emphasized that this is just one particular possible forward model that can be used within the scope of our invention, chosen by way of example due to its relative simplicity. More sophisticated forward models also fall within the scope of the present invention. Such models would take into account, by way of example but not limitation, multiple exposures, various illumination conditions (e.g., offaxis, incoherent), the actual electromagnetics of the light field interacting with the photomask, various types of photomasks besides chrome on glass (e.g., attenuated phase shifting, strong phase shifting, other materials, etc.), the polarization of the light field, the actual properties of the lens (such as aberrations), and a more sophisticated model of the resist (e.g., diffusion within the resist), such as a variable threshold model, lumped parameter model, or a fully three dimensional first principles model.

[0093]
Because the inverse algorithm requires many iterations of the forward algorithm, the latter preferably is implemented efficiently. However, as a general rule, more sophisticated models are likely to run slower than simpler models. One embodiment of the present invention compensates for this difference in model speed by beginning with a simpler model and then gradually introducing more sophisticated models as the process converges, thereby postponing the full complexity until the last iterations. In an alternative embodiment, switching between different models at different time steps obtains an averaging effect. For example, this represents an efficient way to explore the space of errorparameters. Other variations will be apparent to one of ordinary skill in the art.

[0094]
In one embodiment, the Hamiltonian compares the pattern resulting from the forward model with the target pattern in order to determine the figure of merit. For example, an L
_{2}norm may be calculated:
H(ψ(
x,y))=
F(ψ(
x,y))−
T(
x,y)
wherein T(x,y) indicates the target pattern. The L
_{2}norm is indicative of the area of the nonoverlapping regions of the two patterns. This metric approaches zero as the two patterns converge. Other examples of determining a figure of merit are as follows:

 1. Other Norms. These might include a cubic or other polynomial functions of the differences;
 2. Levelset differences. By representing a resulting pattern as a levelset function one can calculate the distance between the boundaries, integrated over the length of the boundaries;
 3. Local variations. Different parts of the image may have different degrees of importance when it comes to variations from the target pattern. For example, gates generally need to be much more accurately printed than interconnects. In one embodiment, a weighting function assigns more weight to nonoverlapping areas in the portions of the design having higher accuracy requirements. A related approach gives priority to a measure of distances between curve, or to other metrics; or
 4. Semantics. Certain types of errors are considered more significant than other types. For example, within small tolerances, variations from the target pattern are irrelevant, whereas variations outside some tolerances are fatal, taking into account the intent of the design and not just the geometry of the design. In one embodiment, use local weightings to account for errors. As an example, consider a gate which must be printed within specific tolerances. Then the weighting factor becomes large for points outside the tolerances. Within tolerances, the weighting factor would be smaller, and optionally still nonzero (so that the algorithm still prefers designs that are closer to the target). Other ways of incorporating design semantics into the merit function will be apparent to one of ordinary skill in the art.

[0099]
Optionally, an adjustment is provided to the Hamiltonian according to empirical measurements of an actual fabrication process.

[0000]
Output

[0100]
The flow chart shown in FIG. 5 ends with an output of the resulting contours, representing a mask suitable for one of the potential photolithography applications and conforming to the specifications and constraints specified in a suitable “pattern I/O format.”

[0101]
Other outputs in addition to the photomask pattern corresponding to the optimized contours are contemplated. In one embodiment, the final value of the Hamiltonian function is output to indicate the merit of the solution, optionally to be interpreted as a probability estimate that the resulting process will print within specification.

[0000]
Generalizations

[0102]
Foregoing discussion frequently considers a single levelset function representing contours on a single mask, the interior of those contours corresponding to chrome regions, and the exterior corresponding to glass regions. However, in many cases it will be desirable to find contours separating multiple types of regions on the same mask, for example, chrome, glass, and phaseshifted glass, and/or either alternatively or simultaneously to find contours corresponding to boundaries of regions on multiple masks, to be used in a multiple exposure process. Both generalization fall within the teachings of our invention.

[0103]
To allow for multiple masks it suffices to simultaneously optimize multiple levelset functions, an algorithm for which follows directly from above discussions: each levelset function timeevolves according to an equation analogous to (2), except that the terms on the right hand side now depend upon multiple levelset functions, instead of just on one.

[0104]
One can easily allow for multiple types of regions in a similar manner to the way in which one handles multiple masks, i.e., with multiple levelset functions. However, with multipletypes of regions on the same mask, one must prevent regions from overlapping. Consider an example where glass regions correspond to those areas in which the first levelset function is positive, phaseshifted regions correspond to those areas in which the second levelset function is positive, and chrome regions correspond to those areas in which both levelset functions are negative. Prohibiting the possibility for the same region to be both clear glass and phaseshifted glass, add a constraint which prevents both levelsets from being positive in the same area, for example by adding a “penalty” term to the Hamiltonian, the penalty term taking on a very large value whenever the two levelsets overlap. Thus, as the system timeevolves, the contours move so as to remain nonoverlapping. It should be apparent that this concept can be extended in a trivial manner to more than two level sets and more than three types of regions. Alternatively, one can allow both levelsets to evolve freely, and assign precedence to one of them, e.g., if both level sets are positive, define the region as clear glass. Other means of representing multiple regions (also called “multiphase flow” in the literature) will be apparent to those skilled in the art, and fall within the scope of our invention.

[0105]
Similarly, while the foregoing discussion refers typically to a mask consisting of only chrome and glass regions, these types of regions should not be construed to limit the applicability of the present invention, which is useful for any number of different types of regions. By way of example (but not limitation), phaseshifted regions, regions covered with materials other than chrome (e.g., in an attenuated phase shifting mask), and halftone regions would all be within the teachings of the present invention.

[0106]
In still another embodiment, a levelset function can be used to represent the pattern of the illumination optics; as in the foregoing discussion of multiple masks, this levelset function can be optimized simultaneously with those representing one or more photomasks. In yet another embodiment, various parameters of the Hamiltonian can be optimized simultaneously with one or more levelset functions, in an analogous manner.

[0107]
Accordingly, while there have been shown and described above various alternative embodiments of systems and methods of operation for the purpose of enabling a person of ordinary skill in the art to make and use the invention, it should be appreciated that the invention is not limited thereto. Accordingly, any modifications, variations or equivalent arrangements within the scope of the attached claims should be considered to be within the scope of the invention.

[0108]
In addition, the foregoing description of the principles of our invention is by way of illustration only and not by way of limitation. For example, although several illustrative embodiments of methodologies in accordance with the principles of our invention have been shown and described, other alternative embodiments are possible and would be clear to one skilled in the art upon an understanding of the principles of our invention. For example, several alternatives have been described for various steps described in this specification. It should be understood that one alternative is not disjoint from another alternative and that combinations of the alternatives may be employed in practicing the subject matter of the claims of this disclosure. Certainly the principles of our invention have utility apart from making photomasks for integrated circuits, some of which we have already mentioned. Accordingly, the scope of our invention is to be limited only by the appended claims.

[0109]
Foregoing described embodiments of the invention are provided as illustrations and descriptions. They are not intended to limit the invention to precise form described. In particular, it is contemplated that functional implementation of invention described herein may be implemented equivalently in hardware, software, firmware, and/or other available functional components or building blocks. Other variations and embodiments are possible in light of above teachings, and it is thus intended that the scope of invention not be limited by this Detailed Description, but rather by claims following.